Thank you Øystein for that obituary and summary of Joseph's work. I was very saddened to hear the news even though my personal contact with him was only through the mailing list.
Gnubg made it's most exciting leagues in playing strength while Joseph was driving that aspect forward. I was aware of some of his innovations but far from all. You amused me with your description of his brusqeuness on the forum; he certainly did not suffer fools gladly, not even imprecision. Thank you, Ian On Fri, 25 Apr 2025, 19:10 Alessandro Scotti, <[email protected]> wrote: > It's very sad to hear this news. I never had the chance to meet Joseph in > person, but I knew of him through his remarkable contributions to computer > backgammon, and I deeply admired his work. > > I'm truly sorry for this loss. My sincere condolences to his family and > loved ones. > > On Fri, Apr 25, 2025 at 12:46 PM Øystein Schønning-Johansen < > [email protected]> wrote: > >> It is with a real sorrow and heavy heart that I write this message. Today >> I became aware of the tragic passing of Joseph Heled. His wife, Edna, has >> informed us through Hans-Jürgen Schäfer the following: >> >> "I have to tell you that Joseph is no longer with us. A month ago he had >> a fatal cardiac arrest while coming back from sports. Sudden, unexpected. >> We are still in shock." >> >> I strongly feel that this is a really big loss to the backgammon >> community and the community of backgammon programmers. Joseph was the >> inventor of so many brilliant ideas that made it into the GNU Backgammon >> code base. He was the main genius behind the true strength of GNU >> Backgammon. The neural networks he trained more 25 years ago are still the >> best backgammon evaluator systems around. This is due to the ingenious >> training method he developed back then. He also found some of the fantastic >> smart methods to speed up evaluation such that GNU Backgammon gained more >> or less perfection. Let me list some of his clever ideas: >> >> - The training method of supervised learning on a fixed dataset from >> rollouts. >> - The benchmark method where move selection is the measured loss. >> - The splitting of neural networks into position classes for race, >> contact and crashed. >> - The introduction of smaller pruning neural networks to find move >> candidates in deeper plies. >> - The table lookup of the exponential function which boosts the >> sigmoid function. >> - The trick of saving the inputs and the linear part of the first >> layer from a previous position evaluation - such that the nn inputs >> become mostly zeros - zeros are skipped such that the nn evaluation goes >> faster. (This might be my personal favorite.) >> >> ... and many more. GNU Backgammon would not have been the same thing >> without his insight and brilliance. >> >> In my communication with Joseph, he was really scientific. He always >> checked his results before stating any conclusion and I respected his >> findings as the ground truth. His insight - even when it came to something >> seamingly simple like race positions - he had analysed things deeper than >> anyone else. Our communications could be mail threads of nearly a hundred >> messages. >> >> He could also be harsh and direct - especially here on this mailing list >> if someone comes with a loose claim or something semi-scientific. In that >> sense - he was not afraid of heated arguments. >> >> He also wrote a computer implementation of the Royal Game of Ur, which is >> considered a precursor to backgammon - and like backgammon - a game that >> combines both skill and luck. https://github.com/jheled/royalur >> >> In the later years he has studied a lot of backgammon and published some >> interesting (and very theoretical) videos on youtube about backgammon race >> doubles and backgammon rating systems. >> https://www.youtube.com/@josephheled-pepster >> >> I strongly believe we have lost one of the best contributors to >> computational backgammon. Thank you, Joseph, and rest in peace. And my >> deepest condolences to Edna and other family members. >> >> -Øystein >> >
